What is a software factory?
Everyone's saying it in 2026. Here's what people mean by it.
The term is old. It used to mean Microsoft's model-driven tools, and later military delivery pipelines like the Air Force's Kessel Run. Now it means you hand an AI system a spec, and it builds and ships working software. People write the spec and check the result.
Most descriptions come down to four parts. Tap through them.
Anatomy of an agentic software factory
Tap each stage. Stage 04 feeds back into stage 01.
What the human does
You stop writing code. You write a spec clear enough to build from, and you decide whether to trust what comes back. Teams get there in steps. First agents only read code, then they open pull requests, then they merge on their own within limits. The five levels of agent autonomy break this down.
How far along it is
Some companies say their agents merge hundreds of pull requests a week. Some of that is true. But most teams get stuck on whether they can trust the output, and review has become the slow part. That's why code review is the bottleneck.
Why the idea still helps
Full autonomy may be years away. Thinking in these four parts still makes you ask useful questions. Who owns the spec? How do agents working in parallel stay out of each other's way? What has to pass before code ships? The rest of this series looks at each part, starting with the spec and the review.